Chip giant Nvidia has launched the third generation of its Nemotron family of artificial intelligence models, datasets, and engineering libraries under an open-source license. The announcement, made on Monday, December 15, 2025, marks a strategic push by Nvidia to expand its influence from AI hardware into the software and model layers. The immediate release is the Nemotron 3 Nano model, with larger Super and Ultra versions scheduled for the first half of 2026.
The Nemotron 3 Nano model boasts significant performance improvements, delivering four times the throughput of its predecessor and reducing reasoning token generation by up to 60%. It employs a hybrid latent mixture-of-experts (MoE) architecture, which has become an industry standard for top open-source models. Its context window expands to one million tokens, making it suitable for long-form documents and complex, multi-step projects. The larger Super and Ultra models leverage Nvidia's proprietary 4-bit NVFP4 training format on its Blackwell hardware to slash memory requirements and training time.
All model weights, a synthetic pretraining corpus of nearly 10 trillion tokens, and detailed training recipes are being released publicly on platforms like GitHub and Hugging Face under the NVIDIA Open Model License. Accompanying the release are open-source libraries for training, reinforcement learning, and safety validation (NeMo Gym, NeMo RL, NeMo Evaluator).
Kari Briski, Nvidia’s Vice President of generative AI software, stated the open release responds to enterprise demand for transparency, particularly in regulated industries like healthcare, finance, and defense that require auditable, on-premises alternatives. "Many of our enterprise customers cannot deploy certain models or build their business on models with opaque source codes," Briski explained, emphasizing the company's commitment to providing a reliable, inspectable model.
The move comes amid a shifting competitive landscape. Chinese tech firms like DeepSeek, Moonshot AI, and Alibaba are gaining ground with their own open-source models, with companies like Airbnb adopting Alibaba's Qwen model. Concurrently, Meta Platforms appears to be retreating from its open-source Llama strategy, potentially leaving Nvidia as a primary American open-source provider. This strategic positioning is further amplified by U.S. government security concerns and bans on certain Chinese AI models, as well as an ongoing anti-monopoly investigation by China into Nvidia's 2020 acquisition of Mellanox.
